TL;DR
Researchers at JPL developed RAPID (Random Access Picture Digitizer), a specialized hardware-memory interface that allowed 1970s-era minicomputers to process video data in near real-time. By decoupling sequential video signals from the computer's random-access needs, they achieved a 10–20x speedup in scene analysis, enabling robots to track objects and perform 3D measurements in seconds rather than minutes.
What Problem Does This Paper Address?
In the mid-1970s, the "video bottleneck" was a major hurdle for robotics. Standard television cameras (EIA RS-170) output data as a continuous, sequential stream of scan lines. However, computer vision algorithms—like edge detection or region growing—need to jump around an image (random access) to follow shapes or compare pixels.
Before RAPID, researchers often had to save image frames to slow disk files and then read them back, which was agonizingly slow. For a robot intended for planetary exploration or underwater assembly, waiting several seconds just to "see" a single object was unacceptable. The authors sought to create a system that could digitize video, store it in a dedicated memory, and allow a computer to pluck out any pixel at any time without interrupting the live video feed.
Methods
The core of the paper is the RAPID system architecture, which acts as a high-speed bridge between two solid-state CID (Charge Injection Device) cameras and a General Automation SPC-16/85 minicomputer.
Key technical features include:
- Double-Buffering Hardware: To handle the speed mismatch between the incoming video (280 ns per pixel) and the computer's memory cycle, RAPID uses two 8-byte buffers. While one buffer is being filled by the camera, the other is written to the main 256x256 byte memory.
- Time-Slicing Access: The system is engineered so that out of every 2.24 microseconds, 1.54 microseconds are reserved specifically for the computer to request data. This allows the computer to access any pixel in 5 microseconds without "flickering" or interrupting the video synchronization.
- Image Analysis Pipeline: The authors implemented a multi-step software process:
- Transformation: Reducing the image to 1/16th size while performing edge enhancement.
- Region Growing: Identifying clusters of pixels that represent objects.
- 3D Stereo Matching: Using two cameras to calculate the "depth" (Z-axis) of an object's centroid by finding the intersection of two rays.
Datasets
While this paper doesn't use modern standardized datasets (like ImageNet), the authors evaluated RAPID across several real-world applications:
- Viking (Mars) Imagery: Processing photos from the Mars landers.
- Robotic Manipulation: Real-time tracking for a robot arm.
- Medical/Biological: Studying facial changes post-surgery and measuring astronaut body volume changes in microgravity.
- Maze Puzzles: Used as a benchmark for scene segmentation.
Key Insights and Ideas
- Hardware-Software Co-design: The "aha moment" was realizing that the computer didn't need constant access to memory. By mathematically timing the gaps between video buffer writes, they created a "window" for the computer to sneak in and grab data.
- Resolution vs. Stability: The authors intentionally chose GE cameras with lower resolution (188x244) than older tube-based cameras because the solid-state CID chips offered better "image linearity and stability." In robotics, knowing exactly where a pixel is in physical space is more important than having a pretty, high-resolution picture.
- The Power of Reduction: By sampling the image at 1/4 resolution and then partitioning it into 2x2 cells for averaging, they could perform complex edge detection much faster, only returning to "full resolution" once a specific object of interest was identified.
Key Results
- Massive Speedups: The entire scene analysis sequence (detecting an object, finding its orientation, and calculating 3D coordinates) took roughly 5 seconds. This was 10 to 20 times faster than previous methods.
- Efficient Stereo Vision: Acquiring grey-level data for 3D correlation took only 0.035 seconds (4% of the total task time) using RAPID, compared to 5 seconds when accessing the same data from a disk file.
- Real-Time Tracking: The system could track moving objects at speeds of 50 to 75 pixels per second, updating the object's position every 1/30th of a second (standard video frame rate).
Limitations and Future Work
- Memory Constraints: The system was limited to a 256x256 byte memory (64 KB), which was significant for the time but small by modern standards.
- Resolution Loss: The authors admit that moving from silicon target vidicons to CID cameras resulted in a loss of resolution, requiring more frequent camera movement to capture details.
- Hardware Cost: While the authors claim RAPID was cheaper than commercial units of the time, it still required specialized TTL logic and custom hardware interfaces that were not easily scalable.
- Future Directions: The paper hints at the need for higher-resolution devices, noting that companies like RCA and Fairchild were beginning to offer sensors with better specs.
Summary updated